A European fertility organization is calling for international limitations on the number of children a single sperm donor can contribute to. This concern stems from experiences where individuals discover they have many unknown half-siblings due to anonymous donations. The push advocates for legal boundaries to manage this aspect of reproductive science. Such limits aim to address the implications for donor-conceived people.
Ondas has acquired DZYNE Technologies for $875.8 million. This acquisition aims to significantly expand Ondas' defense capabilities. DZYNE specializes in several advanced defense technologies. These include long-endurance autonomous aircraft systems. Additionally, DZYNE focuses on counter-drone systems and autonomous effects.
AI agents power the "SceneSmith" system, enabling the creation of realistic virtual playgrounds. These digital environments model real-world settings such as kitchens, hotels, and living rooms. Within these simulated spaces, robots can practice performing daily chores. This process allows for the generation of crucial training data necessary for robot development. Ultimately, the system facilitates practical skill acquisition for robotics through immersive simulation.
LinkForge evolved from a Blender extension to address challenges in robot description authoring formats like URDF and XACRO. The project introduces a programmable Intermediate Representation (IR) designed to improve this process. This IR enables programmatic composition of robot models and validates kinematic/physical properties before export. It aims to function as a tool-agnostic layer preceding existing ROS workflows, rather than replacing them entirely. Authors are seeking community feedback on the need for such an intermediate representation in current robot description pipelines.
OpenAMRobot v0.0.1 is an MIT-licensed, open-source mobile robotics platform developed for ROS 2 Jazzy. It is designed to support education, research, and rapid prototyping in areas like affordable manufacturing and last-mile logistics. Key features of the current release include autonomous navigation, SLAM and map creation, simulation support, and autonomous docking. The platform also provides a web-based operator interface and tools for ROS 2 interaction. Developers are encouraged to provide feedback on architecture, documentation, and potential future improvements.
RoboShield is a low-latency, out-of-band real-time RTPS security watchdog written in Rust. It intercepts and parses raw DDS/RTPS packets from UDP captures to detect anomalies on the network. The system validates each packet's GUID against a configurable whitelist and monitors topic message rates for flood attacks. Unlike solutions requiring node modification like SROS2, RoboShield plugs directly into the network, imposing no changes on existing nodes. Benchmarks show an average processing time of approximately 1.36 ฮผs per packet on x86 hardware.
EgoSteer is a full-stack system designed to enable steerable dexterous manipulation using egocentric human videos. It tackles the lack of suitable demonstration data by integrating EgoSmith, a new pipeline generating 9.6K hours of high-quality pre-training material from in-the-wild videos. The architecture features a unified robot stack for teleoperation and human corrections. After language-guided pre-training on this data, the system undergoes robot post-training and DAgger refinement. Evaluations show EgoSteer robustly executes over 40 diverse tasks, demonstrating generalization and few-shot adaptation to complex goals like box folding.
Addresses safety challenges for robotic manipulators in unstructured environments facing unforeseen obstacles. A novel framework incorporates an environment module for real-time obstacle detection, 3D localization, and geometry estimation, producing a semantic report. Central coordination manages system communication and updates tools like memory and collision scenes. The planning module dynamically selects the best motion planner (e.g., RRTConnect) based on the current goal. This approach refines generated trajectories to guarantee safe, collision-free task execution, demonstrated robustly in Gazebo Classic.
Developing high-level decision making for autonomous driving requires handling dynamically sized inputs due to varying traffic. Existing methods either neglect explicit traffic interactions (like DeepSet) or suffer from high computational complexity when modeling them using attention mechanisms. DecisionPerceiver is proposed, adapting the Perceiver IO architecture by projecting dynamic agent features into a fixed-size latent space via tunable latent queries. This approach improves scalability for handling larger networks while retaining interaction awareness. Evaluation across multiple driving scenarios confirms consistent performance gains and successful generalization even as vehicle density increases.
Accurate trajectory prediction for autonomous driving requires modeling complex, context-dependent agent interactions beyond purely data-driven methods. The paper proposes SWIFT (Small-World Interaction Framework), a unified model that merges small-world networks with traffic flow theory to incorporate structural priors. SWIFT employs a Small-World Interaction Network to capture both local and global dependencies, alongside a Flow Regime Encoder for scene adaptation. Additionally, a multi-relational graph module enhances interaction reasoning by explicitly encoding agent relationships. Experiments on nuScenes, MoCAD, and NGSIM demonstrate that SWIFT surpasses strong baselines in accuracy, generalization, and robustness across varied traffic conditions.
Developing robust autonomous driving systems requires synthesizing and simulating safety-critical scenarios (SCS). Existing methods struggle with maintaining data fidelity when editing agent states or scene elements in existing environments. The proposed OmniSCS system addresses this by generating photorealistic, high-physical-fidelity SCS within synthetic settings for closed-loop testing. It features a module that reconstructs agents and backgrounds while preserving appearance during editing, alongside an SCS Synthesis module for object insertion and trajectory manipulation. Testing on nuScenes, Waymo, and KITTI datasets confirms OmniSCS surpasses current methods in edited scene fidelity and supports real-time validation of self-driving algorithms.
Developing reliable safety validation for autonomous driving necessitates overcoming the limitations of expensive real-world testing. This paper proposes a risk-field enhanced closed-loop digital twin framework to address this gap. The framework integrates physical data acquisition with virtual reconstruction, scenario generation, and algorithm evaluation. A novel "driving risk field" is introduced as a unified representation quantifying various risks around the ego vehicle. Implementing this risk structure allows for targeted, interpretable safety validation, though performance relies on model fidelity and sim-to-real transfer accuracy.
Navigation is crucial for autonomous systems, but current methods often lack real-world robustness. Vision-and-Language Navigation (VLN) integrates language understanding with visual perception for data-driven navigation solutions. This survey systematically reviews VLN by classifying state-of-the-art methods across action and model paradigms. Furthermore, the authors conducted a real-world evaluation on ten scenes, revealing a significant performance drop from simulation to physical deployment. Specifically, while one RGB-only method failed substantially in reality (61% to 22%), a hierarchical framework showed better robustness (51% success rate).